HiFOR: Hierarchical Foresight with Action-Guided World Grounding for Vision-Language-Action Model
Jointly predicts short-term states, semantic milestones and goal states across horizons, then lets the action query retrieve only the relevant parts of that imagined future.跨时间尺度联合预测短期状态、语义里程碑与目标状态，再由动作查询从中检索相关的未来表征。
BibTeX
@misc{ding2026hifor,
title = {{HiFOR: Hierarchical Foresight with Action-Guided World Grounding for Vision-Language-Action Model}},
author = {Rongxing Ding and Hongyu Qu and Kehong Wu and Xudong Wang and Fan Zhang and Xinguang Xiang and Xiangbo Shu},
year = {2026},
eprint = {2610.40373},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.40373},
}